|Publication number||US8238250 B2|
|Application number||US 12/580,253|
|Publication date||Aug 7, 2012|
|Filing date||Oct 16, 2009|
|Priority date||Oct 16, 2009|
|Also published as||US20110090789|
|Publication number||12580253, 580253, US 8238250 B2, US 8238250B2, US-B2-8238250, US8238250 B2, US8238250B2|
|Inventors||Hei Tao Fung|
|Original Assignee||Hei Tao Fung|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (34), Referenced by (5), Classifications (18)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application related to data communications and more particularly to load balancing on link aggregation in a data communications network.
Network devices such as Ethernet switches are interconnected by communication links for receiving and transmitting data packets through physical ports. In order to provide redundancy or larger aggregate bandwidth, some network devices bundle multiple communication links to form link aggregations (LAGs) with their peering network devices. See
In many typical implementations, load sharing is statically configured. For example, packet distribution is based on an algorithm that selects a port based on addresses and session information: source address, destination address, or both source and destination addresses. Packets with the same addresses and session information are always sent to the same port in the LAG to prevent out-of-order packet delivery. Static load balancing algorithms do not take into account the amount of traffic assigned to each port and variation in traffic over time, and they therefore results in suboptimal utilization of the link bandwidth.
Some dynamic load balancing algorithms for LAG have been published. They primarily focus on the idea of calculating hash values based on the packets' addresses and session information and mapping the hash values to physical ports based on the measurements of the traffic load on the physical ports. The weakness of said algorithms is that changing the mappings of hash values to physical ports affects all packet flows with the same hash values, and said algorithms fail to address the impact on preventing out-of-order packet delivery when a large number of packet flows are momentarily assigned to different egress ports. Also, said dynamic load balancing algorithms do not deal with the quality of service (QoS) requirements of packet flows.
A method for QoS-aware dynamic load balancing of packet flows on link aggregation is disclosed. Said method relies on a static load balancing algorithm until near congestion, and in near-congestion condition overrides the default decisions from static load balancing algorithm based on traffic load on physical ports and QoS requirements on a per-flow basis.
In our preferred embodiment, said static load balancing algorithm comprises: calculating a hash value based on the quintuple: destination IP address, source IP address, IP protocol number, destination port number, and source port number; and assigning the egress port based on said hash value. We shall refer to the decisions from said static load balancing algorithm as default port assignments. We shall also use the quintuples to uniquely identify packet flows.
Said near-congestion condition is detected by observing the utilization of the transmit queues of the physical ports in link aggregations. When the transmit queues are full or nearly full, the default port assignments are to be overridden, and the current packet flow is assigned to the physical port that is least congested in the LAG. Also, a flow database is maintained. Each flow entry in said flow database comprises the quintuple of the packet flow, the selected egress port, the selected transmit queue, and the age of the flow entry.
When a packet is to be transmitted over a LAG, the flow database is consulted first and looked up by the quintuple of the packet. When there is a current matching flow entry, the egress port in the flow entry overrides the default port assignment. When there is no current matching flow entry, the default port assignment is used when the default egress port is not in near-congestion condition.
Whether a flow entry is current or not is determined by the age of the flow entry and the current counter value. In our preferred embodiment, a counter and a tag are maintained per transmit queue per physical port. Said counter advances by one unit when a packet, marked by said tag and enqueued in the transmit queue, is transmitted. At any instance, there is at most one packet in a transmit queue marked by said tag. When the packet marked with the tag is transmitted, the tag is free to be attached to the next packet being enqueued in the transmit queue, and the corresponding counter is advanced. When the counter value is significantly different from the age of the flow entry, the flow entry is considered stale and replaceable.
The present disclosure will be understood more fully from the detailed description that follows and from the accompanying drawings, which however, should not be taken to limit the disclosed subject matter to the specific embodiments shown, but are for explanation and understanding only.
A method for QoS-aware dynamic load balancing of packet flows on link aggregation (LAG) is disclosed. Said method relies on a static load balancing algorithm until near congestion, and in near-congestion condition overrides the default decisions from static load balancing algorithm based on traffic load on physical ports and QoS requirements on a per-flow basis.
Static load balancing algorithms are adequate when congestion is far from sight. Taking advantage of that premise, a device that uses the method disclosed herein can dedicate less resource by focusing on dynamic load balancing in near-congestion condition.
In our preferred embodiment, the static load balancing algorithm comprises: calculating a hash value based on a quintuple of the packet consisting of destination IP address, source IP address, IP protocol number, destination port number, and source port number; and assigning the egress port based on said hash value. We shall refer to the decisions from said static load balancing algorithm as default port assignments. We shall also use the quintuples to uniquely identify packet flows. All packets belong to the same packet flow if they have the same quintuple.
As an example, the hash value H is the 16-bit CRC value using the quintuple. One of the P physical ports in the LAG, where P is a positive integer, is selected to be the default egress port by taking modulo P on the hash value.
In our preferred embodiment, we assume that the device using our method disclosed herein comprises a plurality of transmit queues per physical port, and the transmit queues correspond to different scheduling priorities. For example, there is one high-priority transmit queue, one medium-priority transmit queue, and one low-priority transmit queue for each physical port. A transmit queue is a FIFO for scheduling packets for transmissions. Each packet to be transmitted must first be enqueued to one of the transmit queues of the egress port and scheduled to be transmitted on the physical link.
Near-congestion or congestion condition is detected by monitoring the utilization of the transmit queues of the physical ports in the LAGs. When a transmit queue is 100 percent utilized or full, congestion has occurred, and no more packet can be enqueued in the transmit queue. When a transmit queue is more than T-percent utilized, where T is arbitrarily chosen between 0 and 100, the near-congestion condition has occurred. It is reasonable to choose a very high T value. For example, T is 99.
In near-congestion or congestion condition, the default port assignment should be overridden, and the packet being enqueued and its corresponding packet flow are assigned to the least congested physical port in the LAG that offers the targeted scheduling priority. If all transmit queues that correspond to the targeted scheduling priority of all physical ports in the LAG are full and if reducing the targeted scheduling priority of the packet flow under that circumstance is considered desirable, the packet being enqueued can be assigned to a lower scheduling priority transmit queue of one of the physical ports, bearing in mind that out-of-order delivery might occur. If all transmit queues of all ports in the LAG that can offer the targeted scheduling priority of the packet flow are full, the current packet must be discarded.
In order to memorize which packet flows have their default port assignments overridden, a flow database is maintained in the device. In our preferred embodiment, each flow entry in said flow database comprises the quintuple of the packet flow, the selected egress port, the selected transmit queue, and the age of the flow entry. Refer to
When a packet is going through the load balancing decision process, the flow database is consulted first and looked up by the quintuple of the packet. When there is a current matching flow entry, the egress port in the flow entry overrides the default port assignment. When there is no current matching flow entry, the default port assignment is used when the default egress port is not in near-congestion condition. Refer to
In our preferred embodiment, a counter and a tag are maintained per transmit queue per physical port. Whether a flow entry is current or not is determined by the age of the flow entry and the counter value. Said counter advances by one unit when a packet, marked by said tag and enqueued in the transmit queue, is transmitted. At any time, there is at most one packet in a transmit queue marked by said tag. Marking packet with said tag can be implemented as setting a flag in a packet descriptor, modifying a field in the packet, or appending a field to the packet. That is just a way to differentiate a packet from others. When the packet marked with the tag is transmitted, the next packet being enqueued in the transmit queue is marked with the tag, and the corresponding counter is advanced by one. When a flow entry is inserted into the flow database, the age of the flow entry stores the counter value corresponding to the transmit queue of the egress port assigned, and the transmit queue of the flow entry stores the transmit queue assigned. When the counter value is the same as the age of the flow entry or larger than the age of the flow entry by one, the flow entry is considered current; otherwise, the flow entry is considered stale and replaceable. When the counter is advanced through its upper limit, the counter is allowed to be wrapped around. Because of the wrap-around, some flow entries may be evaluated as current even though they are actually stale. Such side-effect is harmless, only that some entries that would have been replaceable are tied up temporarily. Such side-effect can be removed by removing the stale flow entries in the flow database from time to time.
The purpose of marking packet with said tag and using a counter to count the marking of packet with said tag is for detecting when previous packets of a packet flow has finished transmission at the current moment and therefore new packets of the packet flow can be assigned to different ports in the LAG without the risk of out-of-order delivery for the packet flow.
In the example of
In our preferred embodiment, the flow database is implemented as a hash table. Refer to
The embodiments described above are illustrative examples and it should not be construed that the present invention is limited to these particular embodiments. Thus, various changes and modifications may be effected by one skilled in the art without departing from the spirit or scope of the invention as defined in the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5533021 *||Feb 3, 1995||Jul 2, 1996||International Business Machines Corporation||Apparatus and method for segmentation and time synchronization of the transmission of multimedia data|
|US6363077||Feb 12, 1999||Mar 26, 2002||Broadcom Corporation||Load balancing in link aggregation and trunking|
|US6498781||Aug 13, 1999||Dec 24, 2002||International Business Machines Corporation||Self-tuning link aggregation system|
|US6510135 *||Dec 18, 1998||Jan 21, 2003||Nortel Networks Limited||Flow-level demultiplexing within routers|
|US6512742||Dec 9, 1998||Jan 28, 2003||International Business Machines Corporation||System for data transmission across a link aggregation|
|US6535504||Jan 15, 1999||Mar 18, 2003||Brian Weatherred Johnson||Link aggregation path selection method|
|US6614758||Jan 10, 2002||Sep 2, 2003||Broadcom Corp||Load balancing in link aggregation and trunking|
|US6952401||Mar 17, 2000||Oct 4, 2005||Broadcom Corporation||Method for load balancing in a network switch|
|US6970420||Mar 17, 1999||Nov 29, 2005||Nortel Networks Limited||Method and apparatus for preserving frame ordering across aggregated links supporting a plurality of quality of service levels|
|US6977892||Dec 17, 1998||Dec 20, 2005||Nortel Networks Limited||Method and apparatus for preserving flow order across links of a multi link trunk|
|US6981054 *||Jul 18, 2000||Dec 27, 2005||Advanced Micro Devices, Inc.||Flow control arrangement in a network switch based on priority traffic|
|US6993027||Mar 17, 2000||Jan 31, 2006||Broadcom Corporation||Method for sending a switch indicator to avoid out-of-ordering of frames in a network switch|
|US7145869 *||Mar 17, 2000||Dec 5, 2006||Broadcom Corporation||Method for avoiding out-of-ordering of frames in a network switch|
|US7382728 *||May 30, 2002||Jun 3, 2008||Via Technologies, Inc.||Networking switching apparatus and method for congestion control|
|US7545814 *||Feb 28, 2002||Jun 9, 2009||Nokia Corporation||Method and system for dynamic remapping of packets for a router|
|US7668103 *||Dec 1, 2005||Feb 23, 2010||Marvell International Ltd.||Inter-device flow control|
|US7898959 *||Jun 30, 2008||Mar 1, 2011||Marvell Israel (Misl) Ltd.||Method for weighted load-balancing among network interfaces|
|US7936770 *||Mar 8, 2006||May 3, 2011||Enterasys Networks, Inc.||Method and apparatus of virtual class of service and logical queue representation through network traffic distribution over multiple port interfaces|
|US20040015599 *||Apr 14, 2003||Jan 22, 2004||Trinh Man D.||Network processor architecture|
|US20050232274 *||Jun 14, 2005||Oct 20, 2005||Broadcom Corporation||Method for load balancing in a network switch|
|US20060098573 *||Nov 8, 2005||May 11, 2006||Beer John C||System and method for the virtual aggregation of network links|
|US20060198381||Mar 4, 2005||Sep 7, 2006||Anusankar Elangovan||Method and apparatus for efficient load distribution on link aggregations|
|US20070041321 *||Dec 22, 2005||Feb 22, 2007||Fujitsu Limited||Network switch apparatus that avoids congestion at link-aggregated physical port|
|US20070147238 *||Jan 29, 2007||Jun 28, 2007||Broadcom Corporation||Method for managing congestion in a network switch|
|US20080181103 *||Jan 29, 2007||Jul 31, 2008||Fulcrum Microsystems Inc.||Traffic distribution techniques|
|US20080291826 *||May 24, 2007||Nov 27, 2008||Harris Stratex Networks Operating Corporation||Dynamic Load Balancing for Layer-2 Link Aggregation|
|US20080298236||Jun 1, 2007||Dec 4, 2008||Cisco Technology, Inc.||Dynamic link aggregation|
|US20090190580||Jul 30, 2009||Paramesh Kailasam||Method and apparatus for Link aggregation using links having different link speeds|
|US20090207857 *||Feb 13, 2009||Aug 20, 2009||Alexander Jr Cedell A||User-specified key creation from attributes independent of encapsulation type|
|US20100091780 *||Oct 8, 2009||Apr 15, 2010||Brocade Communication Systems, Inc.||Frame traffic balancing across trunk groups|
|US20100106866 *||Oct 24, 2008||Apr 29, 2010||Juniper Networks, Inc.||Flow consistent dynamic load balancing|
|US20110051602 *||Mar 3, 2011||Broadcom Corporation||Dynamic load balancing|
|US20110051735 *||Mar 3, 2011||Broadcom Corporation||Dynamic load balancing using virtual link credit accounting|
|US20110090789 *||Oct 16, 2009||Apr 21, 2011||Hei Tao Fung||QoS-aware Flow-based Dynamic Load Balancing for Link Aggregation|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US8358597 *||Oct 1, 2009||Jan 22, 2013||Hei Tao Fung||Method for building scalable Ethernet switch network and huge Ethernet switch|
|US20110080855 *||Apr 7, 2011||Hei Tao Fung||Method for Building Scalable Ethernet Switch Network and Huge Ethernet Switch|
|US20130215743 *||Feb 21, 2013||Aug 22, 2013||Marvell World Trade Ltd.||Network devices with time aware medium access controller|
|US20140115167 *||Dec 23, 2013||Apr 24, 2014||Marvell Israel (M.I.S.L) Ltd.||Load Balancing Hash Computation for Networks|
|US20140204761 *||Oct 7, 2013||Jul 24, 2014||Brocade Communications Systems, Inc.||Optimizing traffic flows via dynamic routing protocol modifications when using server virtualization with dynamic routing|
|U.S. Classification||370/237, 370/235, 370/230|
|International Classification||H04J3/14, G01R31/08, H04J1/16, H04L12/26, G08C15/00, G06F11/00, H04L1/00|
|Cooperative Classification||H04L47/125, H04L47/11, H04L47/6215, H04L41/5025, H04L47/2408|
|European Classification||H04L47/11, H04L41/50B2, H04L47/12B|